Adaptive martingale approximations (Q734947): Difference between revisions

Consider a probability space \((\Omega, {\mathcal A}, P)\) and the associated Hilbert space \(L^2(\Omega, {\mathcal A}, P)\). An orthonormal system of functions \((u_k; k \geq 0)\) defined on \(\Omega\) is called an \textit{H-system} if and only if for any \(X \in L^2(\Omega, {\mathcal A}, P)\) \[ X_{{\mathcal A}_n} \equiv {\mathbf E}(X|u_0,u_1, \dots, u_n)= \sum_{k=0}^n (X, u_k)u_k,\quad \text{for\;all}\;n \geq 0. \] This paper studies this class of orthonormal systems and their potential to perform lossy compression. From a theoretical point of view they are interesting as the sequence of approximations is a martingale. Moreover, for a given collection of random variables \({\mathcal X} = (X_1 \dots X_d)\), the authors show how to construct the system \((u_k)\) adaptively to obtain efficient approximations. Here the word \textit{adapted} is used in two different ways. First, the system \((u_k)\) will be constructed in an optimal way using the set \(\mathcal X\) (and hence adapted to \(\mathcal X\)) and, second, the \(u_k\) are simple functions which are adapted (according to measure theory) to the sigma algebras \(\sigma(u_0,\dots,u_n)\) which form a natural filtration in the sense that the filtration generates \(\sigma({\mathcal X})\). The main contribution of this paper is in the construction, via the greedy splitting algorithm, of H--systems adapted to a collection of random variables and in showing the usefulness of these approximations in a concrete setting.